Abstract: The performance of speech recognition system will be significantly deteriorated because of the mismatches between training and testing conditions. This paper addresses the problem and proposes an environment adaptation algorithm to adapt the mean vectors of HMM. The algorithm can reduce the performance deterioration of the speech recognition system caused by the mismatches. Firstly, we build a binary tree by Gaussian similarity analysis (GSA) and then adaptively adjust the class number according to the data. In each class, we adapt the HMM using nonlinear transform approximated by piecewise linear regression. Rather than using maximum likelihood estimation (MLE) in estimating the transformation parameters, we propose using maximum a posteriori (MAP) as the estimation criterion. The proposed algorithm, called GAS-MAPNT, has been evaluated on a Chinese digit recognition experiment based on continuous density HMM. The test shows that the proposed algorithm is efficient and superior to other algorithms with Gaussian similarity analysis, such as maximum a posteriori linear regression (MAPLR) algorithm and maximum likelihood linear regression (MLLR) algorithm.

Key words: speech recognition, nonlinear transform, Gaussian similarity analysis, maximum a posteriori